fix: budget tracking, system prompt loading, and reasoning model support

- Track tool_calls_used in delegation loop budget (was always 0)
- Add max_tool_calls check to can_continue() and budget_fraction_remaining
- Load system prompt from agent config path before convention path fallback
- Add robust JSON extraction from LLM responses with text wrapping
- Support reasoning models (Nemotron) that return content in reasoning field
- Add E2E integration test notebook (07) covering all AWP modalities
- Update notebook defaults to nvidia/nemotron-3-super-120b-a12b:free

Tested end-to-end with OpenRouter + Nemotron 3 Super: 9/9 tests pass
(LLM connectivity, tool calling, DAG A0/A1, delegation loop A2/A3,
AgentWorkflow with DataFrame/tool creation A4).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: veegee82

examples/jupyter/tutorials/04_dag_workflow_e2e.ipynb

@@ -44,7 +44,7 @@
     "\n",
     "# --- OpenRouter (cloud) ---------------------------------------\n",
     "OPENROUTER_API_KEY = \"\"                 # paste your key or set env var OPENROUTER_API_KEY\n",
-    "OPENROUTER_MODEL = \"openrouter/google/gemini-2.5-flash\"\n",
+    "OPENROUTER_MODEL = \"nvidia/nemotron-3-super-120b-a12b:free\"\n",
     "\n",
     "# --- Custom OpenAI-compatible API -----------------------------\n",
     "CUSTOM_API_KEY = \"\"                     # paste your key or set env var LLM_API_KEY\n",
@@ -1055,4 +1055,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 5
-}
+}

examples/jupyter/tutorials/05_delegation_loop_e2e.ipynb

@@ -58,7 +58,7 @@
     "\n",
     "# --- OpenRouter (cloud) ---------------------------------------\n",
     "OPENROUTER_API_KEY = \"\"  # paste your key or set env var OPENROUTER_API_KEY\n",
-    "OPENROUTER_MODEL = \"openrouter/google/gemini-2.5-flash\"\n",
+    "OPENROUTER_MODEL = \"nvidia/nemotron-3-super-120b-a12b:free\"\n",
     "\n",
     "# --- Custom OpenAI-compatible API -----------------------------\n",
     "CUSTOM_API_KEY = \"\"                     # paste your key or set env var LLM_API_KEY\n",
@@ -1460,4 +1460,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 5
-}
+}

examples/jupyter/tutorials/06_full_showcase.ipynb

@@ -57,7 +57,7 @@
     "\n",
     "# --- OpenRouter (cloud) ---------------------------------------\n",
     "OPENROUTER_API_KEY = \"\"  # paste your key or set env var OPENROUTER_API_KEY\n",
-    "OPENROUTER_MODEL = \"openrouter/google/gemini-2.5-flash\"\n",
+    "OPENROUTER_MODEL = \"nvidia/nemotron-3-super-120b-a12b:free\"\n",
     "\n",
     "# --- Custom OpenAI-compatible API -----------------------------\n",
     "CUSTOM_API_KEY = \"\"\n",
@@ -11587,7 +11587,7 @@
     "  result = AgentWorkflow(\n",
     "      inputs={\"data\": df, \"config\": {...}},\n",
     "      task=\"Analyze trends and create a report\",\n",
-    "      model=\"openrouter/google/gemini-2.5-flash\",\n",
+    "      model=\"nvidia/nemotron-3-super-120b-a12b:free\",\n",
     "      max_loops=5,\n",
     "      code_mode=True,\n",
     "      tool_creation=True,\n",
@@ -11742,4 +11742,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 4
-}
+}

examples/jupyter/tutorials/07_e2e_integration_test.ipynb

@@ -157,14 +157,55 @@ def _run_simple(self, messages: list[dict], **kwargs: Any) -> dict[str, Any]:
         try:
             return self._llm.chat_json(messages, **kwargs)
         except json.JSONDecodeError:
-            # Retry with explicit JSON instruction
+            # Try to extract JSON from within the response text
             text = self._llm.chat_text(messages, **kwargs)
+            extracted = self._extract_json(text)
+            if extracted is not None:
+                return extracted
             logger.warning("Agent %s: non-JSON response, wrapping", self.name)
-            return {"result": text, "confidence": 0.3}
+            return {"result": text, "confidence": 0.5}
         except Exception as exc:
             logger.error("Agent %s LLM error: %s", self.name, exc)
             return {"error": str(exc), "confidence": 0.0}
 
+    @staticmethod
+    def _extract_json(text: str) -> dict[str, Any] | None:
+        """Try to extract a JSON object from text that may contain extra content."""
+        if not text:
+            return None
+        # Try the full text first
+        stripped = text.strip()
+        # Strip markdown fences
+        if stripped.startswith("```"):
+            lines = stripped.split("\n")
+            lines = [ln for ln in lines if not ln.strip().startswith("```")]
+            stripped = "\n".join(lines).strip()
+        try:
+            result = json.loads(stripped)
+            if isinstance(result, dict):
+                return result
+        except (json.JSONDecodeError, ValueError):
+            pass
+        # Find first { ... } block
+        start = text.find("{")
+        if start == -1:
+            return None
+        depth = 0
+        for i in range(start, len(text)):
+            if text[i] == "{":
+                depth += 1
+            elif text[i] == "}":
+                depth -= 1
+                if depth == 0:
+                    try:
+                        result = json.loads(text[start : i + 1])
+                        if isinstance(result, dict):
+                            return result
+                    except (json.JSONDecodeError, ValueError):
+                        pass
+                    break
+        return None
+
     def _run_with_tools(
         self, messages: list[dict], tool_defs: list[dict], **kwargs: Any
     ) -> dict[str, Any]:
@@ -193,17 +234,11 @@ def _run_with_tools(
             if not content:
                 return {"result": "No response", "confidence": 0.0}
 
-            # Parse as JSON
-            cleaned = content.strip()
-            if cleaned.startswith("```"):
-                lines = cleaned.split("\n")
-                lines = [line for line in lines if not line.strip().startswith("```")]
-                cleaned = "\n".join(lines).strip()
-
-            try:
-                return json.loads(cleaned)
-            except json.JSONDecodeError:
-                return {"result": content, "confidence": 0.3}
+            # Parse as JSON (try robust extraction)
+            extracted = self._extract_json(content)
+            if extracted is not None:
+                return extracted
+            return {"result": content, "confidence": 0.5}
 
         except Exception as exc:
             logger.error("Agent %s tool loop error: %s", self.name, exc)
@@ -215,8 +250,12 @@ def _build_system_prompt(self) -> str:
         """Assemble system prompt: base + skills + memory."""
         parts: list[str] = []
 
-        # 1. Base system prompt
-        base = self._load_file("workflow/instructions/SYSTEM_PROMPT.md")
+        # 1. Base system prompt — try config path first, then convention path
+        base = ""
+        if self._config.prompt.system:
+            base = self._load_file(self._config.prompt.system)
+        if not base:
+            base = self._load_file("workflow/instructions/SYSTEM_PROMPT.md")
         if base:
             parts.append(base)
         else:

reference/python/src/awp/runtime/agent.py

@@ -96,6 +96,8 @@ def budget_fraction_remaining(self) -> float:
             1 - (self.workers_spawned / max(self.max_total_workers, 1)),
             1 - (self.wall_time_elapsed / max(self.max_wall_time, 1)),
         ]
+        if self.max_tool_calls:
+            fractions.append(1 - (self.tool_calls_used / self.max_tool_calls))
         return max(0.0, min(fractions))
 
     def can_continue(self) -> tuple[bool, str]:
@@ -106,6 +108,8 @@ def can_continue(self) -> tuple[bool, str]:
             return False, "max_total_workers reached"
         if self.wall_time_elapsed >= self.max_wall_time:
             return False, "max_wall_time exceeded"
+        if self.max_tool_calls and self.tool_calls_used >= self.max_tool_calls:
+            return False, "max_tool_calls reached"
         return True, "ok"
 
     def to_dict(self) -> dict[str, Any]:
@@ -1348,6 +1352,7 @@ def _tracking_call(name: str, arguments: dict) -> dict:
                             "result": result,
                         }
                     )
+                    self._budget.tool_calls_used += 1
                     return result
 
                 final_msg = llm.chat_with_tools(

reference/python/src/awp/runtime/delegation_loop_runner.py

@@ -381,12 +381,24 @@ def chat_text(
         messages: list[dict[str, Any]],
         **kwargs: Any,
     ) -> str:
-        """Return just the assistant's text content."""
+        """Return just the assistant's text content.
+
+        Handles reasoning models that may put output in the ``reasoning``
+        field when ``content`` is null or empty.
+        """
         data = self.chat(messages, **kwargs)
         choices = data.get("choices", [])
         if not choices:
             return ""
-        return choices[0].get("message", {}).get("content", "") or ""
+        msg = choices[0].get("message", {})
+        content = msg.get("content", "") or ""
+        if not content:
+            # Reasoning models (e.g. Nemotron) may return content in
+            # the reasoning field when max_tokens is constrained
+            reasoning = msg.get("reasoning", "")
+            if reasoning:
+                content = reasoning
+        return content
 
     def chat_json(
         self,